Comparisons of sequentially presented vibrotactile frequencies have been extensively studied using electrophysiological recordings in nonhuman primates. Although neural signatures for working memory aspects of such tasks were recently also identified in human oscillatory EEG activity, homologue correlates of the comparison process are yet unknown. Here, we recorded EEG activity while participants decided which of two sequentially presented vibrotactile stimuli had a higher frequency. Because choices in this type of task are known to be systematically biased by the time-order effect, we applied Bayesian modeling to account for individual choice behavior. Using model-based EEG analysis, we found that upper beta band amplitude (∼20–30 Hz) was modulated by participants' choices. The modulation emerged ∼750 msec before a behavioral response was given and was source-localized to premotor areas.Importantly, the choice-dependent modulation of beta band amplitude was invariant to different motor response mappings and reflected the categorical outcome of the subjective comparison between the two frequencies. Consistently, this pattern was evident for both correct and incorrect trials, indicating that the beta band amplitude mirrors the internal representation of the comparison outcome. Our data complement previous findings in nonhuman primates and corroborate that the beta band activity in premotor areas reflects the categorical outcome of a sensory comparison prior to translation into an effector-specific motor command.
Over the last decades, studies in nonhuman primates have identified neuronal mechanisms that underlie memory-based perceptual decisions in the somatosensory domain (reviewed in Romo & de Lafuente, 2013). In their seminal work, Romo and colleagues used a vibrotactile two-alternative forced-choice (2AFC) task. Monkeys were trained to decide whether the second of two sequentially presented vibrotactile stimuli (with frequency f2) had a lower or higher frequency than the first one (with frequency f1). Thus, f1 served as the (variable) reference value against which f2 had to be compared. Electrophysiological recordings in several parietal and frontal brain areas revealed a cascade of neuronal processes involved in this task (see Romo & de Lafuente, 2013, for details): (i) during presentation of the tactile stimuli, firing rates in primary and secondary somatosensory cortices (SI and SII) scaled monotonically with the frequency of the stimuli (Hernández, Zainos, & Romo, 2000); (ii) firing rates of neurons in the PFC were parametrically modulated by the values of f1 during the entire retention interval of the task (Romo, Brody, Hernández, & Lemus, 1999; see also Barak, Tsodyks, & Romo, 2010); (iii) crucially, correlates of the comparison process underlying the perceptual decision were evident in firing rates of medial and ventral premotor cortex (mPMC and vPMC, respectively; Romo, Hernández, & Zainos, 2004; Hernández, Zainos, & Romo, 2002). More specifically, neuronal activity in these areas reflected the (to-be-evaluated) signed frequency difference between f1 and f2. Consequently, the premotor cortex could be regarded as a candidate area in which the comparison between the sample stimulus (i.e., f2) and the memory trace of f1 is computed.
More recently, recordings of monkey local field potentials (LFPs) showed that the central role of the premotor cortex in sequential frequency comparisons is also expressed in terms of oscillatory activity. The spectral power of beta band LFP oscillations (∼18–26 Hz) recorded from the mPMC mirrored the categorical decisions of the monkeys (Haegens et al., 2011). Specifically, the beta band power was significantly increased when monkeys indicated that f2 was higher than f1, both for correct and incorrect choices. Hence, a decision-related signal in the vibrotactile 2AFC task is not only encoded in the firing rates of single neurons (cf. Romo et al., 2004; Hernández et al., 2002) but also in synchronized ensemble activity on the neural population level, corroborating a central role of neural oscillations in information processing (e.g., see Siegel, Donner, & Engel, 2012).
Along these lines, recent studies in humans found oscillations in a similar frequency band (∼20–30 Hz) to reflect working memory (WM)-related processes during the retention interval of vibrotactile 2AFC tasks (Spitzer & Blankenburg, 2011; Spitzer, Wacker, & Blankenburg, 2010). In particular, EEG recordings showed that upper beta power recorded over PFC scaled parametrically with the frequency that had to be kept in WM (Spitzer & Blankenburg, 2011; Spitzer et al., 2010). However, the decision-related aspects of the vibrotactile 2AFC task have not been investigated in human EEG recordings yet.
In this study, we aimed to close the gap between the results from invasive electrophysiological studies of decision processes in monkeys, on the one hand, and the available findings of WM correlates in oscillatory human EEG data, on the other. In particular, we asked whether human oscillatory EEG signals may encode the decision-relevant quantity (i.e., the difference between f2 and f1) that reflects the comparison of both stimuli (cf. Haegens et al., 2011). To answer this question, we optimized the sequential frequency comparison paradigm to allow for an artifact-free analysis of EEG data during the decision period of the task. Moreover, we applied Bayesian modeling to account for a well-known systematic order effect in choice behavior that is commonly observed in sequential 2AFC comparisons: the so-called time-order error/effect (TOE; cf. Sanchez, 2014; Karim, Harris, Morley, & Breakspear, 2012; Ashourian & Loewenstein, 2011; Preuschhof, Schubert, Villringer, & Heekeren, 2010; Woodrow, 1935; Fechner, 1860). The TOE refers to the finding that, in 2AFC tasks, participants tend to compare the second stimulus against a weighted average of the mean of the whole stimulus set and the first stimulus, instead of the first stimulus alone (cf. Karim et al., 2012). Our Bayesian model yielded individual estimates of these subjectively perceived frequency differences for each stimulus pair as a proxy for the internal representation of the trial-specific f1-versus-f2 comparison. Using time–frequency (TF) analysis, we studied decision-related oscillatory EEG signals by examining in which frequency bands spectral amplitude was correlated with the estimated subjective frequency differences. We found such decision-related signals in the form of categorical EEG amplitude modulations in the upper beta band over premotor areas, in notable agreement with recent invasive recordings in monkeys.
Twenty-four healthy, right-handed volunteers (21–35 years; 14 women) participated in the experiment after giving written informed consent. The study was approved by the local ethics committee at the Freie Universität Berlin. Six participants (three men, three women) were excluded from the analysis because of chance-level behavioral performance (<60% correct answers, five participants) or excessive EEG artifacts (one participant).
Stimuli and Behavioral Task
Suprathreshold vibrotactile stimuli were applied to the left index finger using a piezoelectric Braille stimulator (QuaeroSys Medical Devices, Schotten, Germany). The 16 pins (4 × 4 matrix, consistent peak amplitude ∼0.46 mm) of the stimulator's display were driven by a sinusoidal carrier signal (fixed at 133 Hz) that was amplitude-modulated by a lower-frequency sinusoid (varied between 12 and 32 Hz). The resulting stimulation created the sensation of a tactile “flutter” stimulus (Romo & Salinas, 2003; Talbot, Darian-Smith, Kornhuber, & Mountcastle, 1968) at the modulation frequency (i.e., 12–32 Hz), whereas the spectrum of the physical driving signal was confined to frequencies above 100 Hz (e.g., Tobimatsu, Zhang, & Kato, 1999). Thus, the risk of physical artifacts in the EEG analysis range (<100 Hz) was minimized. To mask the sound of the stimulator, white noise of ∼80 dB was played during the whole experiment via loudspeakers that were placed below a TFT monitor in front of the participant (e.g., Spitzer & Blankenburg, 2011; Spitzer et al., 2010). Throughout the task, a fixation cross was displayed at the center of the TFT monitor. On each trial, two brief vibrotactile stimuli (with frequencies f1 and f2) were presented for 250 msec each, separated by a retention interval of 1000 msec (Figure 1A). Such short-lived stimulations evoke only transient signals in EEG data and hence facilitate a clean analysis of the decision period immediately after the second stimulus. The values of f1 were randomly chosen from 16, 20, 24, or 28 Hz; f2 could differ from f1 by ±2 or 4 Hz (Figure 1B). Thus, participants could not predict the difference f2 − f1 based on f1. After presentation of the second stimulus, participants indicated whether f2 or f1 was higher by pressing one of two buttons with their right index or middle finger, respectively. Importantly, the response assignment of the buttons was reversed for half of the participants, such that the mapping of choices onto specific button presses (which might have been associated with specific motor preparatory signals) was fully counterbalanced across participants. Twenty milliseconds after each response, performance feedback was provided in form of two “plus” or “minus” signs indicating correct or incorrect responses, respectively, presented to the left and to the right of the fixation cross for 190 msec. The next trial started after a variable intertrial interval (1500–2000 msec). Participants completed seven blocks of 160 f1-versus-f2 comparisons (each block lasted ∼12 min), for a total of 1120 trials. Before the experiment started, participants performed ∼50 practice trials.
To estimate the individually perceived subjective frequency differences that can account for TOE-like biases in sequential comparison tasks, we fitted a Bayesian inference model to the behavioral data. The model we used was adopted from earlier work (cf. Sanchez, 2014; Ashourian & Loewenstein, 2011).
To assess the model's goodness-of-fit, we compared the individual model fits with a “null” model in which decisions were only based on the physical stimulus differences (f2 − f1). That is, the posteriors of f1 and f2 are both directly represented by their respective likelihood function, and only and b was to be estimated. Bayes factors (BFs) were computed for each participant to quantify the goodness-of-fit of the subjective decision model relative to the “null” model while accounting for differences in model complexity (e.g., see Kass & Raftery, 1995).
In addition, we evaluated the predictions of the subjective decision model on an independent test set to control empirically for overfitting. In particular, we randomly divided the trials of each participant condition-wise into a training set and into a test set. Parameters of the model were estimated on the training set and then applied for fitting the test data (cf. Figure 4B–D).
EEG Recording and Analysis
EEG was recorded from 64 electrodes (ActiveTwo, BioSemi, Amsterdam, The Netherlands) positioned in an elastic cap according to the extended 10–20 system. Four additional electrodes were used to register horizontal and vertical eye movements. Individual electrode locations for each participant were obtained prior to the experiment using a stereotactic electrode-positioning system (Zebris Medical GmbH, Isny, Germany). The EEG data were digitized at 2048 Hz, offline down-sampled to 512 Hz, high- and low-pass filtered (with cutoff frequencies of 0.5 and 48 Hz, respectively), and re-referenced to a common average montage. Eye blinks were corrected using adaptive spatial filtering based on individual calibration data (for details, see Ille, Berg, & Scherg, 2002). In addition, trials with signal amplitudes exceeding a threshold of 80 mV (before low-pass filtering) were excluded from further investigations (12.2% of trials on average). The analyses were done in MATLAB (The MathWorks, Natick, MA) using the SPM8 toolbox (Wellcome Department of Cognitive Neurology, London, UK; www.fil.ion.ucl.ac.uk/spm), including the FieldTrip toolbox for EEG/MEG data (Radboud University Nijmegen, Donders Institute; fieldtrip.fcdonders.nl). Unless stated otherwise, only trials with correct choices were used for analysis.
The artifact-free EEG data were segmented into epochs from −2500 to 1000 msec relative to the time of the button press to examine the decision period of the task (i.e., response-locked analysis). TF representations of spectral power between 4 and 48 Hz (in steps of 2 Hz) were computed every 50 msec by applying a Morlet wavelet transformation with a sliding window of 7 cycles length (i.e., TF bin = 50 msec × 2 Hz). Exploratory analysis of higher-frequency bands (up to 100 Hz), using a multitapered Fourier transformation with three Slepian tapers and a sliding window length of 400 msec, yielded no significant effects.
For the main analysis of decision-relevant signals (Figure 4), the TF transformation was applied to single-trial response-locked data, yielding a measure of ongoing “whole” power. Power changes in overall “evoked” (i.e., phase-locked) and “induced” (i.e., non-phase-locked) activity were analyzed on stimulus locked data (−2250 to 2250 msec relative to f2 onset). Evoked power associated with each stimulus pair (f1, f2) was assessed by applying the TF transform to the average (time domain) waveform of the corresponding trials. Induced power was computed by applying the TF transform to single trials from which the ERP, associated with the respective stimulus pair (f1, f2), was subtracted beforehand.
The response-locked single-trial TF data were square root transformed (yielding spectral amplitudes) to approximate normally distributed data (see Kiebel, Tallon-Baudry, & Friston, 2005). To decrease intersubject variability, TF data were smoothed with a 3 Hz × 300 msec FWHM Gaussian kernel (e.g., Litvak et al., 2011; Kilner, Kiebel, & Friston, 2005). For individual participants, we regressed the spectral amplitude in each TF bin of each channel on the zero-centered estimates of subjective frequency differences across trials: We created a vector (across single trials) of subjective frequency differences (i.e., f2 − f1′) for each participant, subtracted the mean value, and used the vector as a predictor for between-trial variations of spectral amplitude in each TF bin. Hence, we estimated TF maps that quantified the linear relation between the individual subjective frequency differences (i.e., the results from our Bayesian model) and the spectral amplitude in each TF bin. To identify times, frequencies, and channels for which this linear relationship was significantly different from zero, we used cluster-based permutation testing (Maris & Oostenveld, 2007). We compared the summary statistics of the observed data (one-sample t test across participants in each TF bin) with a distribution of summary statistics obtained from 500 randomly sign-flipped permutations. A cluster was defined as a group of adjacent TF bins that all exceeded a cluster-defining threshold of pthreshold < .001 (uncorrected). Clusters that exceeded a family-wise error (FWE)-corrected threshold of pcluster < .05 (corrected for time, frequency, and channels) were considered to be statistically significant.
On the basis of the distribution of subjective frequency difference values across participants, we binned the values into seven levels, such that for every level at least one stimulus pair (i.e., subjective frequency difference value) from each participant was available (i.e., [−0.66 to −0.33]; [−0.33 to −0.18]; [−0.18 to −0.09]; [−0.09 to 0]; [0 to 0.09]; [0.09 to 0.17]; [0.17 to 0.4]). Individual EEG data were grouped according to these levels to assess grand mean time courses and to localize the cortical source of the choice-modulated beta band signal as follows.
The cortical sources of amplitude modulations observed on the scalp level were localized using the 3-D source reconstruction routines provided by SPM8 (Friston, Henson, Phillips, & Mattout, 2006). On the basis of the individually recorded electrode positions for each participant (substituted by default 10–20 locations for two participants because of technical difficulties), a forward model was constructed using a 8196-point cortical mesh of distributed dipoles perpendicular to the cortical surface of a template brain (cf. Friston et al., 2008). The lead field of the forward model was computed using the three-shell boundary elements method EEG head model available in SPM8. Multiple sparse priors (Friston et al., 2008) under group constraints (Litvak & Friston, 2008) were used to invert the forward model. For each condition, the results of model inversion were summarized in a 3-D image that reflected spectral source amplitude in the TF window of interest. Relevant contrasts of these 3-D images served as an estimate for subject-specific source locations and were used for group level statistical analysis (see Litvak et al., 2011). Anatomical reference for source estimates was established on the basis of the SPM anatomy toolbox (Eickhoff et al., 2005) where possible.
Choice-modulated beta band activity was localized using the preprocessed response-locked EEG data (i.e., in the time domain). Additionally, the data were band-pass filtered in the frequency range of the TF cluster identified on the scalp level (±1 Hz to ensure that no information is lost at the cluster borders; see Figure 4A). Before inverting the forward model, single trials of each participant were grouped according to the seven levels of subjective frequency differences (see Time Courses). The 3-D images summarizing each condition were computed over a representative TF window (20–30 Hz; −750 to −350 msec from button press). To identify cortical sources in which beta band amplitude was modulated by subjective frequency differences, the 3-D images were weighted by a contrast vector analogously to the sensor space analysis. Source estimates were statistically analyzed on the group level using conventional t tests and displayed at a threshold of p < .01 (uncorrected).
On average, participants made correct choices on 74.0% of all stimulus pairs. For detailed analysis, we performed a within-subject ANOVA with the factors Difficulty (±4 Hz vs. ±2 Hz stimulus difference) and Sign (positive vs. negative stimulus difference) on proportions of correct responses (PCRs), using a logit-transform to account for nonnormality of the residuals. The analysis revealed significant main effects of the factors Difficulty (p < .001) and Sign (p = .014) and a significant interaction of the two factors (p = .001). As expected, a larger proportion of trials were judged correctly when the (physical) f2 − f1 frequency difference was ±4 Hz (81.0% correct) compared with trials where the difference was only ±2 Hz (67.0%; p < .001; paired t test; see difficulty effect, Table 1). We also observed more correct responses for positive (77.8% correct) compared with negative frequency differences (70.3%; p = .014 paired t test; see sign effect, Table 1), which indicates an overall response bias toward “f2 > f1” choices (mean criterion shift: 0.0881; p = .0076; one-sample t test).
|.||Frequency Difference of Stimuli (f2 − f1) in Hz .||Difficulty Effect .||Sign Effect .|
|−4 .||−2 .||2 .||4 .|
|PCR (%)||75.9 ± 5.8||64.8 ± 4.9||69.3 ± 3.5||86.2 ± 4.1||− (p < .001)*||− (p = .014)*|
|RT correct (msec)||842.2 ± 58.1||867.7 ± 54.7||818.0 ± 65.0||777.9 ± 57.7||32.9 ± 14.1 (p < .001)*||57.0 ± 28.8 (p < .001)*|
|RT incorrect (msec)||917.6 ± 78.4||889.2 ± 75.8||933.1 ± 69.6||971.6 ± 81.0||−33.4 ± 30.7 (p = .034)*||−49.0 ± 37.7 (p = .014)*|
|.||Frequency Difference of Stimuli (f2 − f1) in Hz .||Difficulty Effect .||Sign Effect .|
|−4 .||−2 .||2 .||4 .|
|PCR (%)||75.9 ± 5.8||64.8 ± 4.9||69.3 ± 3.5||86.2 ± 4.1||− (p < .001)*||− (p = .014)*|
|RT correct (msec)||842.2 ± 58.1||867.7 ± 54.7||818.0 ± 65.0||777.9 ± 57.7||32.9 ± 14.1 (p < .001)*||57.0 ± 28.8 (p < .001)*|
|RT incorrect (msec)||917.6 ± 78.4||889.2 ± 75.8||933.1 ± 69.6||971.6 ± 81.0||−33.4 ± 30.7 (p = .034)*||−49.0 ± 37.7 (p = .014)*|
PCRs and RTs as a function of the physical frequency difference f2 − f1. Mean values ± 95% confidence interval are shown. Difficulty effect compares easy (±4 Hz) and difficult (±2 Hz) trials in a paired t test. Sign effect compares trials with a positive (2 and 4 Hz) and negative (−4 and −2 Hz) frequency difference in a paired t test. RTs showed significant effects of difficulty and sign for correct and incorrect trials, however, in opposing directions (cf. interactions in ANOVA of RTs). PCRs were logit-transformed before testing, because of nonnormally distributed residuals. Asterisks indicate statistically significant results.
An ANOVA (2 × 2 × 2 repeated-measures design with factors Correct/incorrect, Difficulty, and Sign) of the median RTs showed a significant main effect for the factor Accuracy (p < .001) and two significant interactions (Accuracy × Sign and Accuracy × Difficulty, all ps < .001). More precisely, the median RT with respect to f2 stimulus onset was on average shorter for correct trials (826 msec) than for incorrect trials (927 msec; p < .001; paired t test). The precise pattern of interaction effects in the RT data is detailed in Table 1.
Bayesian Inference Model Describes Individual Choice Behavior
Assuming that the difference f2 − f1 (both frequencies on logarithmic scale according to Weber's law) describes the comparison of vibrotactile frequencies in our task, one would expect PCRs as illustrated in Figure 1B. That is, PCRs for trials with |f2 − f1| = 4 Hz should always be higher than for trials with |f2 − f1| = 2 Hz, independent of the specific frequency values of f1 and f2. However, we observed strong and systematic departures from such response behavior (Figure 1D and E). In particular, the proportion of “f2 > f1” choices increased with increasing f1, whereas the proportion of “f1 > f2” choices increased with decreasing f1 (Figure 1D). Both trends intersected at the mean frequency of the stimulus set. This systematic and symmetric bias reflects the characteristic influence of the TOE on choices in sequential 2AFC comparison tasks (cf. Sanchez, 2014; Preuschhof et al., 2010).
In other words, the observed choice pattern suggests that participants showed a tendency to compare f2 with the mean of the stimulus set (i.e., with a representation of f1 that had regressed to the mean of the stimulus set). As a consequence, the mere difference of the physical magnitudes f2 − f1 is not sufficient to account for the comparison process that drives decisions in our task (Figure 2B vs. C; cf. Hellström, 2003). Hence, we used a Bayesian model (cf. Petzschner et al., 2015; Sanchez, 2014; Ashourian & Loewenstein, 2011) that accounts for the influence of the TOE and yields estimates of the subjectively perceived frequency differences (f2 − f1′) for each stimulus pair and participant (see Bayesian model and Figure 2A). In our model, the posterior distribution of f1 is used to describe the representation of f1 by incorporating a priori knowledge about the stimulus set as well as the actual value of f1. Consequently, the posterior (centered on f1′) is closer to the mean of the stimulus set than the true value of f1 (cf. f1 vs. f1′ in Figure 2A). Simulated choices based on comparing f2 with f1′ (i.e., f2 − f1′) approximated the PCR on the test data set for each participant very well (Figure 2B). Furthermore, the estimated model parameters accounted for individual differences in behavioral measures across participants. The estimated precision of stimulus encoding () was clearly correlated with conventional d′ values (r = .87, p < .001; Figure 2C), whereas the individual influence of the TOE was well described by the participants' precision of prior knowledge (; r = .85, p < .001; Figure 2D). In other words, the higher the influence of prior knowledge on the percept of f1, the larger the influence of the TOE on choices (cf. Karim et al., 2012). As anticipated, also the estimated values of the bias term b were highly correlated with individuals' general response bias toward responding “f2 > f1” or “f1 > f2” independent of any stimulus information (i.e., criterion shift; r = .92, p < .001). Using BFs to quantify the quality of the proposed model (as compared with a “null” model based on the physical differences f2 − f1) provided positive evidence (BF > 3) in favor of the proposed model for each participant (all BFs > 7, strong evidence with BF > 20 for 12 participants; cf. Kass & Raftery, 1995). Furthermore, BFs were also highly correlated with the individual influence of the TOE on choice behavior (r = .96, p < .001), indicating that accounting for the TOE is the reason for the improved model fit.
Stimulus-evoked Activity and Task-induced TF Modulations
At first, we verified the presence of well-documented somatosensory stimulus effects in the EEG recordings. Figure 3A illustrates the TF representation of steady-state-evoked potentials collapsed across representative electrodes (Fz, F2, FC2, FCz, CP6, CP4, P4, P6; see inset, Figure 3A) during stimulus presentation for an exemplary stimulus pair (f1 = 20 Hz; f2 = 24 Hz). As expected, the evoked TF spectrum prominently mirrors the frequency and duration of the presented stimuli (Figure 3A).
The TF representation of the grand-average induced power (Figure 3B) mimicked the typical pattern reported in previous EEG studies of vibrotactile 2AFC tasks (e.g., Spitzer et al., 2010). Results are shown for illustrative electrodes C3, C4, C5, C6, CP3, CP4, CP5, CP6, and POz (see inset, Figure 3B). Throughout the trial, a marked increase in occipital alpha band activity (8–12 Hz) was evident (Figure 3B). Furthermore, we observed a power decrease over bilateral somatosensory areas during/after f1 presentation in the alpha/mu (8–12 Hz) to beta (15–25 Hz) frequency bands (Figure 3B). This power decrease was followed by a rebound (i.e., a recovery with subsequent increase beyond baseline/previous level) of beta band power in electrodes over contralateral (i.e., right hemispheric) somatosensory areas (∼400 to 800 msec after f1; Figure 3B). Finally, in the time interval between f2 stimulus offset and participants' responses, we observed a decrease in beta band power peaking over ipsilateral (left) primary motor cortex (electrode C3; Figure 3B). All effects reported here passed a conservative test on statistical significance (cluster-based permutation test, pthreshold < .001, and pcluster < .05, FWE-corrected).
Beta Band Oscillations in Premotor Areas Encode Choice Independent of Response Mapping
To test if the frequency comparison (f1 vs. f2) was reflected in oscillatory EEG activity, we used the subjectively perceived frequency differences as inferred from our Bayesian model (i.e., f2 − f1′) as regressors for a linear regression analysis of each participant's single-trial TF spectra. The analysis revealed a positive relationship between the upper beta band amplitude (∼20–30 Hz) and the subjective frequency differences in medial–frontal electrodes (FCz, FC2, and C2; inset Figure 4A) well before responses were given (−750 to −350 msec from response; pcluster = .019, FWE-corrected; Figure 4A, dashed rectangle). More specifically, high values of the signed subjective frequency differences (i.e., the decision-relevant quantity f2 − f1′) were associated with high amplitudes, whereas low (i.e., negative) values were reflected by low amplitudes. The scalp topography (Figure 4B) and 3-D source localization (Figure 4C) of the TF cluster suggest that the beta band modulation originated from premotor areas (Brodmann's area 6, peak coordinates in MNI space: 20, −6, 68). Qualitatively very similar results were obtained when using physical frequency differences (i.e., f2 − f1) for the analysis, which is an expected outcome given that subjective and physical frequency difference values were highly correlated (r ≥ .7 for every participant).
In a control analysis, we checked whether the observed modulation of upper beta band amplitude could have been explained by the values of f2 alone: We repeated the main analysis as described above on a subset of data in which only trials were considered with f2 values that could lead to either choice (i.e., “f2 > f1” and “f2 < f1”). With this reduced data set, we found the same modulation of upper beta band amplitude (∼20–30 Hz) consistent in location (FC2) and time (−850 to −500 msec from response; pcluster = .074, FWE-corrected), indicating that the modulation was indeed decision related. Furthermore, we tested whether the overall response bias in behavior might have contributed to the finding that “f2 > f1” choices are associated with a higher upper beta band amplitude. To this end, we computed another control analysis using only data of participants that showed no substantial response bias (absolute value of criterion shift < 0.1; six participants). Despite the reduced sample size, this analysis still revealed the same tendency of “f2 > f1” choices being reflected in a higher beta band amplitude (for correct and incorrect choices).
Figure 4 (D and E) shows the grand mean time courses of upper beta band amplitude for correct and incorrect trials. Importantly, the definition of correct and incorrect was based on the modeled subjective frequency differences, that is, we classified objectively incorrect trials (e.g., “f2 > f1” choice, but f2 < f1) as subjectively correct (and vice versa) if choices were in accordance with the subjectively perceived frequency difference as inferred from Bayesian modeling (e.g., f2 > f1′; 16 participants showed at least one of these swaps). For subjectively correct trials (Figure 4D), we computed the grand mean time courses for each of the seven subjective difference levels as inferred from Bayesian modeling (see Methods). Note that the aforementioned swaps from objectively incorrect to subjectively correct trials (and vice versa) only occurred in the two subjective difference levels bordering zero (the lightest blue and the lightest red in Figure 4D). Taking these swaps into account led to a significant increase in the difference of beta band amplitude between the two affected levels of subjective difference as compared with using the objective definition of correct and incorrect trials (paired t test, p = .019). Overall, beta band amplitude appeared categorically modulated according to the two types of choice, rather than being linearly modulated by the subjective frequency difference values. The black bar in Figure 4D indicates a time window within which pairwise statistical tests of beta band amplitude between choice categories (“f2 > f1” or “f2 < f1”) were significantly different (paired one-sided t test, p < .05, false discovery rate-corrected), whereas none of the pairwise tests reached significance within choices (p > .05, false discovery rate-corrected). For subjectively incorrect trials (i.e., choices that were not in line with the modeled subjectively perceived frequency differences), we computed the grand mean time courses for two classes only (f2 > f1′ and f1′ > f2; Figure 4E) because of insufficient trial numbers for some levels of subjective frequency differences in individual participants. Importantly, a categorical bifurcation of upper beta band amplitude according to choices was also observed for subjectively incorrect trials (black bar: paired one-sided t test, p < .05; Figure 4E). Taken together, the upper beta band activity seems to reflect the internal representation of a subjective quantity (i.e., the subjective comparison outcome) that determines choices (correct and incorrect) in the given task.
Finally, we asked if the choice-related modulation of beta band amplitude depended on a specific mapping of choices onto button presses. Therefore, we divided participants into two groups according to their response mapping (i.e., whether the index or middle finger was used to indicate “f2 > f1”). For each group, we investigated separately whether the medial–frontal beta band amplitude was in the same way choice-modulated as observed for the entire group (i.e., higher beta band activity for “f2 > f1” choices for correct and incorrect trials). On the basis of the results above, we pooled correct and incorrect trials and compared the average beta band amplitude between “f2 > f1” and “f1 > f2” choices for each participant. Figure 5 displays the two group level scalp topographies of the difference in beta band amplitude (between choices) for the previously identified TF cluster (−750 to −350 msec from response, 20–30 Hz). The beta band modulations in the two groups were of the same sign (i.e., higher beta band amplitude for “f2 > f1” choices), showed considerable topographical overlap (cf. white dots in Figure 5A and B span cluster for whole group), and were statistically indistinguishable (independent two-sample t test comparing both scalp topographies revealed no clusters, all ps > pthreshold). In other words, the choice-related modulations of beta band amplitude were not systematically linked to the subsequent execution of a specific motor response associated with either choice.
We investigated oscillatory EEG signatures of perceptual decisions that are based on comparing two sequentially presented vibrotactile frequencies f1 and f2. Medial–frontal upper beta band amplitude (∼20–30 Hz) was modulated by participants' choices, regardless of the specific motor response mapping. In particular, choices of “f2 > f1” were always accompanied by higher beta band amplitude than “f1 > f2” choices. Importantly, these choice-related modulations of oscillatory activity were evident clearly before responses were given and were source-localized to premotor areas. Our findings extend previous studies linking neuronal activity in PMC of nonhuman primates to vibrotactile comparisons (Haegens et al., 2011; Hernández et al., 2002, 2010; Romo et al., 2004). Neuronal firing rates in vPMC and mPMC were shown to be modulated by the signed difference between vibrotactile frequencies (i.e., f2 − f1; Hernández et al., 2002, 2010; Romo et al., 2004). Moreover, Haegens et al. (2011) reported a choice-related signal on the population level, that is, in form of amplitude modulations of beta band oscillations in monkey PMC. Here, for the first time, we extend these findings to human observers with remarkable consistency in terms of quality (higher amplitudes for “f2 > f1” choices independent of accuracy and motor response), putative source (PMC), and frequency (beta band). Because we recorded EEG from the whole scalp, our data additionally suggest (within the scope of EEG) that such modulations of beta band amplitude prior to overt responding might in fact be specific to premotor areas. Notably, Haegens et al. (2011) used a delayed response protocol, whereas our task allowed an immediate response. This difference is likely to explain why Haegens et al. (2011) found the choice-related modulation of beta band amplitude accompanied by a beta peak, whereas we see the same effect on top of an overall beta band desynchronization (i.e., power decrease; most likely because of the preparation of the ensuing button press and propagated via volume conductance from left motor cortex to those electrodes that show the choice-related amplitude modulation).
In contrast to previous studies investigating decision processes in the vibrotactile comparison task, we did not use the physical frequency differences (f2 − f1) to characterize choices in this task. Instead, we modeled subjectively perceived frequency differences (f2 − f1′) for each stimulus pair based on individual behavioral data. Thus, we could account for a characteristic bias in choice behavior commonly observed in 2AFC comparison tasks (cf. TOE; Sanchez, 2014; Ashourian & Loewenstein, 2011), while obtaining a proxy for the putative internal representation of the underlying comparison (i.e., the subjectively perceived frequency difference f2 − f1′). Expanding on classic psychophysical models (cf. Sanchez, 2014; Karim et al., 2012; Hellström, 1979, 1985), subjectively perceived frequency differences in our parsimonious model were defined as the difference between f2 and a weighted average of f1 and the mean of the stimulus set, expressed in terms of Bayesian inference (cf. Petzschner et al., 2015). For each participant, the model based on subjectively perceived frequency differences (f2 − f1′) explained the behavioral data significantly better than a comparable model based only on physical/objective differences (f2 − f1; BF > 7 for all participants). Although both models yielded qualitatively similar results in the EEG analyses, the subjective difference model permitted a considerably more fine-grained scale of individual frequency differences (16 subjective frequency differences vs. 4 physical frequency differences), revealing that the time courses of beta band amplitude clearly separated into only two distinct choice-specific levels. The EEG analysis based on subjectively perceived frequency differences thus indicates more conclusively than an analysis using physical frequency differences that the observed amplitude modulation in the beta band during decision-making is presumably categorical, rather than parametric (monotonic). This suggests that premotor beta band amplitude does not represent the (relative or absolute) decisional evidence per se, but rather the categorical outcome of the internal comparison (see also Haegens et al., 2011). Under this view, additional, potentially more finely graded comparison processes are likely to occur upstream of the large-scale oscillatory signal disclosed in the present analysis.
Notably, the present amplitude modulation in the beta band was inverted for incorrect trials, underpinning the interpretation that this activity corresponds to an internal representation of the subjective decision outcome. In other words, the beta band amplitude reflects whether an observer is about to choose “f2 > f1” or “f1 > f2.” Importantly, the choice outcome was disentangled from specific motor responses in our study, in contrast to studies that exploit lateralized EEG signals in preparation for a motor response to predict decisions (e.g., Polanía, Krajbich, Grueschow, & Ruff, 2014; O'Connell, Dockree, & Kelly, 2012; Schurger, Sitt, & Dehaene, 2012; Donner, Siegel, Fries, & Engel, 2009). In the present work, participants indicated decisions by pressing one of two buttons, always with their right hands, using different fingers. We counterbalanced the response mapping across participants and found the same modulation of beta band amplitude no matter which finger was used to respond. Similarly, Haegens and colleagues showed that the choice-related modulation of beta band amplitude in monkeys was absent when no f1-versus-f2 comparison was required, but a prespecified button was pressed (Haegens et al., 2011; see also Romo et al., 2004; Hernández et al., 2002). Both approaches converge on showing that the beta band amplitude modulations in premotor areas are decision related (i.e., choice-selective) and not merely linked to a (specific) motor response (i.e., effector-selective; see e.g., Polanía et al., 2014; O'Connell et al., 2012; Schurger et al., 2012; Donner et al., 2009). Taken together, upper beta band amplitude in PMC seems to represent subjective choices (i.e., the subjectively categorized outcome of a quantitative comparison) that are not yet expressed in specific motor terms.
Besides the choice-related modulations, we also found typical patterns of sensorimotor beta band oscillations (∼15–25 Hz) that are routinely observed during somatosensory and motor tasks. That is, when a tactile stimulus is presented or anticipated, beta band activity is known to decrease over somatosensory areas and to rebound ∼600 msec afterwards (e.g., Van Ede, de Lange, Jensen, & Maris, 2011; Bauer, Oostenveld, Peeters, & Fries, 2006; Pfurtscheller, 1981; Jasper & Andrews, 1938). In preparation for and during a voluntary hand movement, the same pattern of beta band desynchronization, followed by a rebound, can be observed over contralateral motor areas (e.g., Pfurtscheller, 1981; Jasper & Penfield, 1949). However, it appears unlikely that the choice-related amplitude modulations in the upper beta band we observed over premotor areas are epiphenomena of these classic sensorimotor signals: We confined our analysis to response-locked data to render confounding effects of motor preparation unlikely in the first place. In response-locked data, systematic RT variations should affect beta band amplitude only in form of systematically time-shifted stimulus-locked signals (e.g., a beta band rebound after f2). However, the observed time courses of upper beta band amplitude showed no sign of any time-shifted components (Figure 4D and E). Lastly, we can also rule out that the observed modulations in the upper beta band might be explained by generally higher f2 values in “f2 > f1” choices. When using a subset of data, including only trials in which f2 values could lead to either choice, we still found the same beta band amplitude modulations in electrodes over premotor areas. Taken together, the reported findings are highly unlikely to be the result of a systematic stimulus or response artifact.
In addition to a notable consistency between our results and previous work in nonhuman primates (e.g., Haegens et al., 2011), our findings also connect well with human EEG studies that investigated parametric WM correlates. In vibrotactile comparisons with longer delay periods, Spitzer and colleagues found that upper beta band amplitude (∼20–30 Hz) in PFC was systematically modulated by the to-be-maintained vibrotactile frequency information (Spitzer & Blankenburg, 2011; Spitzer et al., 2010). In particular, during the retention interval of the task, the frequency of the first stimulus (f1) was encoded by the upper beta band amplitude. Further work suggests that upper beta band amplitude might encode analogue WM-related quantity information in a supramodal, generalized fashion (Spitzer, Gloel, Schmidt, & Blankenburg, 2014; Spitzer & Blankenburg, 2012). From this perspective, the present results suggest that the upper beta band amplitude in respective brain areas (PFC and/or PMC) seems to represent task-relevant quantities during the according phases of the vibrotactile comparison task. That is, a detailed representation of absolute quantity during retention (in form of parametric modulations in PFC, see also Barak et al., 2010; Romo et al., 1999) and a categorical representation of the comparison outcome relating to either choice before responding (in form of categorical, choice-dependent modulations in PMC, see also Haegens et al., 2011; Romo et al., 2004; Hernández et al., 2002).
The present finding should be differentiated from a previously reported association of beta band amplitude with the accuracy of a decision (Donner et al., 2007). Donner and colleagues found that in a visual motion detection task beta band amplitude in the dorsal visual pathway was higher for correct trials than for incorrect trials. As discussed by the authors, this finding may relate to the computations that are involved in forming a decision and, in particular, might index the confidence of a decision. In contrast, the results of this study provide evidence for the upper beta band amplitude to represent a quantity on which a perceptual decision is based (see also Siegel, Engel, & Donner, 2011; Donner et al., 2007; deCharms & Zador, 2000). Together with recent studies of WM (see also Spitzer et al., 2014), the present results might suggest beta band activity as a “spectral fingerprint” (see Siegel et al., 2012) of large-scale neural activity involved in the internal evaluation of analogue/quantitative information. However, it remains to be shown in future research how such content representation arises mechanistically in the amplitude of upper beta band oscillations.
To conclude, during vibrotactile frequency comparisons, upper beta band amplitude (∼20–30 Hz) in premotor areas was modulated by the choice of participants, independent of a specific motor response and regardless of the correctness of the choice. The topography, timing, and frequency range of the reported signal are in notable agreement with previous findings in nonhuman primates performing an analogue task. In particular, premotor upper beta band amplitude encoded subjective choices prior to translation into an effector-specific motor command. Hence, we suggest that this signal is an internal representation of the subjective categorical outcome of the comparison underlying perceptual decisions in the vibrotactile comparison task.
This work was supported by the Deutsche Forschungsgemeinschaft (DFG, GRK1589/1). We thank Simon Ludwig and Sebastian Fleck for help with data acquisition and Jakub Limanowski for useful comments on the manuscript. Moreover, we want to thank Gaëtan Sanchez and Jeremie Mattout for providing the code of their model, on which we based the implementation of our model.
Reprint requests should be sent to Jan Herding, Neurocomputation and Neuroimaging Unit, Department of Education and Psychology, Freie Universität Berlin, Habelschwerdter Allee 45, 14195 Berlin, Germany, or via e-mail: email@example.com.
The expression “f1 > f2” refers to a choice, whereas f1 > f2 describes the relation between the physical values of f1 and f2.
These authors contributed equally to the work.